Long context: who reprices the whole request, and who does not
2026/09/02
Feeding a model a long document changes the per-token price at some providers and not at others, and this is easy to miss because the model name does not change.
Five providers here have a threshold[1,2,3,4,7]. OpenAI reprices gpt-6-astra above 272,000 input tokens, to $20 input and $75 output against $10 and $50[1]. GPT-6 Sol and GPT-6 Luna also double their input and cache rates once a prompt passes 272,000 input tokens[8,9]. xAI reprices every Grok model at 200k, so grok-4.6 goes from $2 and $6 to $4 and $12[2].
Google steps Gemini 3.1 Pro Preview at 200k, from $2 and $12 to $4 and $18[3]. Alibaba steps qwen3.7-plus at 256K, from $0.4 to $1.2 on input and from $1.6 to $4.8 on output[4]. Sakana steps fugu-ultra-v2.0 at 272K, from $5 and $30 to $10 and $45[7].
At OpenAI, xAI and Alibaba, crossing the line by one token raises the price of every token before it as well[1,2,4].
OpenAI writes that the long-context rates apply to the full request, and xAI that a request reaching the threshold is billed at the higher rate for all tokens in it[1,2]. Alibaba writes that all tokens in the request are billed at the unit price of the matching tier[4].
Google and Sakana publish the threshold without saying which way it counts[3,7]. Google labels its rates as prompts up to and over 200k tokens, and Sakana labels one column as context over 272K[3,7]. Neither page says whether the tokens below the line are repriced too[3,7].
Claude 4.6 and later include the full 1 million token context at the standard rate[5]. A 900k token request is billed at the same per-token rate as a 9k one, and caching and batch discounts apply across the whole window[5].
Moonshot publishes no threshold for kimi-k3, so its 1,048,576 tokens run at one rate[6]. Sakana's fugu-max-v1.0 is also priced the same at any context length[7].
